A fan blade defect recognition method and system based on multi-feature fusion

Abstract

This invention relates to the field of defect identification technology and discloses a method and system for identifying wind turbine blade defects based on multi-feature fusion. The method includes: acquiring acoustic emission signals and performing noise reduction processing; using time-frequency decomposition to separate damage-related frequency intervals and determine the distribution range of potential damage signals; performing sensitivity calibration on signal intensity and marking suspected defect signal segments; combining historical vibration data and operating load, and determining defect signal intervals through time-series comparative analysis; fusing changes in ambient temperature and wind speed to calculate the matching degree between defects and stress distribution, and obtaining defect impact assessment indicators; when the indicators exceed safety thresholds, generating targeted maintenance instructions and determining repair priorities; and finally generating a dynamic monitoring plan to continuously track defect intervals and obtain long-term defect evolution data. This invention enables high-precision identification of wind turbine blade defects, deep fusion of multi-source features, and intelligent maintenance decision support based on dynamic monitoring.

Description

Technical Field

[0001] This invention relates to the field of defect identification technology, and in particular to a method and system for identifying wind turbine blade defects based on multi-feature fusion. Background Technology

[0002] Currently, with the rapid development of the wind power industry, wind turbine blades, as key load-bearing components of wind turbine units, are directly related to the safe operation and power generation efficiency of the entire unit. Because blades are subjected to complex and variable climatic environments and dynamic loads for extended periods, they are prone to latent defects such as matrix cracking, fiber breakage, and interface debonding. If these defects are not detected and addressed in a timely manner, they may lead to serious safety accidents.

[0003] In existing technologies, defect identification of wind turbine blades often relies on periodic inspections or vibration monitoring, typically employing traditional signal processing methods to analyze the collected acoustic emission signals. However, these methods struggle to effectively separate environmental noise from defect signals under complex operating conditions and lack the ability to perform multi-dimensional fusion analysis of defect characteristics, resulting in low identification accuracy and a high false alarm rate. Furthermore, existing methods fail to fully utilize the multi-source information provided by intelligent sensor arrays, hindering accurate location and dynamic tracking of defect signals and failing to meet the practical needs of wind turbine blade structural health monitoring.

[0004] In summary, existing technologies struggle to effectively distinguish between environmental noise and defect signals under complex operating conditions, resulting in low accuracy in wind turbine blade defect identification. Summary of the Invention

[0005] This invention provides a method and system for identifying wind turbine blade defects based on multi-feature fusion, so as to effectively distinguish between environmental noise and defect signals under complex operating conditions and improve the accuracy of wind turbine blade defect identification. Firstly, to address the aforementioned technical problems, this invention provides a method for identifying wind turbine blade defects based on multi-feature fusion, comprising: Acoustic emission signal data of wind turbine blades are collected, and the acoustic emission signal data is denoised by adaptive filtering to obtain a denoised signal dataset. Based on the denoised signal dataset, time-frequency decomposition is used to separate specific frequency ranges related to damage, based on the pre-acquired signal timestamp records and signal acquisition duration, to determine the distribution range of potential damage signals. The signal intensity variation characteristics within the distribution range are obtained, and the signal intensity variation characteristics are calibrated using the pre-acquired acquisition device sensitivity to obtain the signal energy distribution. If the signal energy distribution exceeds a preset energy threshold, it is marked as a suspected defect signal segment, thus obtaining a set of suspected defect signals. Historical vibration data and operating load records for the time interval corresponding to the suspected defect signal set are obtained, and waveform envelope deviation analysis is performed in combination with preset historical defect labels to identify abnormal fluctuation segments and determine the defect signal interval. Based on the defect signal range, combined with the pre-acquired ambient temperature changes and wind speed fluctuations, a correlation mapping operation is used to calculate the matching quantization value between the defect signal range and the blade stress distribution. The matching quantization value is then weighted to obtain the defect impact assessment index. If the defect impact assessment index exceeds the preset safety threshold, the defect status information is generated and the defect concern level is determined by matching and analyzing the preset status response template with the pre-acquired cumulative blade operation time and the defect signal range data. Based on the defect concern level, combined with the historical vibration data and the changes in ambient temperature, a dynamic monitoring plan is generated to continuously track the defect signal range and obtain long-term defect evolution data.

[0006] Secondly, the present invention provides a wind turbine blade defect identification system based on multi-feature fusion, comprising: The signal acquisition and denoising module is used to acquire acoustic emission signal data of the wind turbine blades and perform denoising processing on the acoustic emission signal data using adaptive filtering to obtain a denoised signal dataset. The time-frequency decomposition module is used to separate specific frequency ranges related to damage based on the denoised signal dataset and the pre-acquired signal timestamp records and signal acquisition duration, thereby determining the distribution range of potential damage signals. The suspected defect marking module is used to acquire the signal intensity change characteristics within the distribution range, calibrate the signal intensity change characteristics using the pre-acquired acquisition device sensitivity, and obtain the signal energy distribution. If the signal energy distribution exceeds a preset energy threshold, it is marked as a suspected defect signal segment, thus obtaining a suspected defect signal set. The defect interval determination module is used to acquire historical vibration data and operating load records of the time interval corresponding to the suspected defect signal set, and perform waveform envelope deviation analysis in combination with preset historical defect labels to identify abnormal fluctuation segments and determine the defect signal interval. The defect impact assessment module is used to calculate the matching quantization value between the defect signal range and the blade stress distribution based on the defect signal range and the pre-acquired ambient temperature change and wind speed fluctuation, and to perform a weighted calculation on the matching quantization value to obtain the defect impact assessment index. The defect concern generation module is used to generate defect status information and determine the defect concern level if the defect impact assessment index exceeds a preset safety threshold, based on a preset status response template and in combination with the pre-acquired cumulative blade operation time and the defect signal range data. The dynamic monitoring module is used to generate a dynamic monitoring plan based on the defect concern level, combined with the historical vibration data and the changes in ambient temperature, to continuously track the defect signal range and obtain long-term defect evolution data.

[0007] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention collects acoustic emission signal data from wind turbine blades and uses adaptive filtering to denoise the data, effectively suppressing environmental background noise and electromagnetic interference, and improving the signal-to-noise ratio and data quality. Combined with the multi-channel layout of the intelligent sensor array, it can fully cover the key structural areas of the blades, providing a highly reliable data foundation for subsequent defect identification and solving the problem of inaccurate data acquisition in complex environments in existing technologies.

[0008] (2) This invention extracts specific frequency ranges related to damage through time-frequency decomposition and calibrates the signal strength in conjunction with equipment sensitivity to accurately identify suspected defect signal segments. This method fully considers the propagation and attenuation characteristics of signals in composite materials, improves the accuracy of defect location, avoids the problem of missing weak defect signals in traditional methods, and enhances the system's ability to perceive early defects.

[0009] (3) This invention accurately determines the defect signal range by performing time-series comparative analysis of suspected defect signals with historical vibration data and operating load records, and combining this with a convolutional neural network model for similarity matching. This process integrates multi-source time-series features, improves the intelligence level and accuracy of defect identification, and solves the problems of single features and low recognition rate in traditional methods.

[0010] (4) This invention further combines ambient temperature change and wind speed fluctuation data, and uses the correlation mapping method to analyze the matching degree between defect signals and blade stress distribution, constructs defect impact assessment indicators, and quantifies the impact of defects on structural safety. By introducing thermodynamic and aerodynamic coupling analysis, the scientificity and practicality of the assessment results are improved, providing a reliable basis for operation and maintenance decisions.

[0011] (5) This invention generates targeted maintenance instructions based on defect impact assessment indicators, and determines the priority of local repairs by combining the cumulative operating time of the blades and the distribution of structural weak points, thereby achieving precise scheduling of maintenance resources. At the same time, it continuously tracks the defect area based on a dynamic monitoring plan, and predicts the evolution trend of defects by combining an LSTM model, providing early warning of potential risks, which significantly improves the intelligence level and operation and maintenance efficiency of blade structural health management. Attached Figure Description

[0012] Figure 1 This is a flowchart illustrating an embodiment of a wind turbine blade defect identification method based on multi-feature fusion provided by the present invention. Figure 2 This is a flowchart illustrating an embodiment of the defect signal range determination provided by the present invention.

[0013] Figure 3 This is a schematic diagram of an embodiment of a wind turbine blade defect identification system based on multi-feature fusion provided by the present invention. Detailed Implementation

[0014] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0015] Reference Figure 1 The first embodiment of the present invention provides a method for identifying wind turbine blade defects based on multi-feature fusion, comprising the following steps: S11, Acoustic emission signal data of wind turbine blades are collected, and the acoustic emission signal data is denoised by adaptive filtering to obtain a denoised signal dataset; S12, Based on the denoised signal dataset, and for the pre-acquired signal timestamp records and signal acquisition duration, time-frequency decomposition operation is used to separate specific frequency ranges related to damage, and the distribution range of potential damage signals is determined. S13, acquire the signal intensity change characteristics within the distribution range, calibrate the signal intensity change characteristics using the pre-acquired acquisition device sensitivity, and obtain the signal energy distribution. If the signal energy distribution exceeds a preset energy threshold, it is marked as a suspected defect signal segment, and a suspected defect signal set is obtained. S14, acquire historical vibration data and operating load records for the time interval corresponding to the suspected defect signal set, perform waveform envelope deviation analysis in combination with preset historical defect labels, identify abnormal fluctuation segments, and determine the defect signal interval; S15. Based on the defect signal range and combined with the pre-acquired ambient temperature changes and wind speed fluctuations, a correlation mapping operation is used to calculate the matching quantization value between the defect signal range and the blade stress distribution. The matching quantization value is then weighted to obtain the defect impact assessment index. S16, If the defect impact assessment index exceeds the preset safety threshold, according to the preset state response template, combined with the pre-acquired cumulative blade operation time and the data of the defect signal range, a matching analysis is performed to generate defect state information and determine the defect attention level. S17. Based on the defect concern level, combined with the historical vibration data and the ambient temperature change, a dynamic monitoring plan is generated to continuously track the defect signal range and obtain long-term defect evolution data.

[0016] In step S11, acoustic emission signal data of the wind turbine blades is collected, and adaptive filtering is used to denoise the acoustic emission signal data to obtain a denoised signal dataset, including: Acoustic emission signals generated by the wind turbine blades during operation are acquired through a sensor array; The acoustic emission signal is discretized according to the sampling frequency to obtain a feature sequence containing waveform characteristics and frequency distribution; Adaptive filtering is used to remove noise interference components present in the feature sequence to obtain the residual signal; If the residual signal retains the complete signal intensity change characteristics, then the residual signal is determined to be a cleaned signal; if it does not retain them, then the residual signal is removed, and the cleaned signal is integrated to obtain a denoised signal dataset.

[0017] First, acoustic emission signals generated by the wind turbine blades during operation are acquired using a sensor array. Specifically, piezoelectric acoustic emission sensors are used in the sensor array. These sensors have the characteristics of wide response bandwidth, high sensitivity, and suitability for acquiring acoustic emission signals from composite material structures. The sensor placement is determined based on the structural characteristics of the wind turbine blades and historical failure cases. Sensors are placed at the blade root bolt connection area, one-third of the blade length from the blade root, two-thirds of the blade length from the blade root, the leading edge and trailing edge of the lightning protection zone at the blade tip, and the main beam cap, forming a multi-channel spatial array to ensure complete coverage of acoustic emission signal acquisition across the entire blade range. The sampling frequency of the acoustic emission signals is determined based on the acoustic emission characteristics of wind turbine blade defects. The acoustic emission signal frequencies generated by typical defects in the composite material of wind turbine blades, such as matrix cracking and fiber breakage, are concentrated in the range of 20kHz-1MHz. According to the Nyquist sampling theorem, the sampling frequency must be higher than twice the highest frequency of the signal. In this embodiment, the sampling frequency is set to 2MHz.

[0018] Subsequently, the acoustic emission signal is discretized according to the sampling frequency to obtain a feature sequence containing waveform characteristics and frequency distribution. It should be noted that the core of the discretization process is to convert the continuous analog acoustic emission signal collected by the sensor into a digital signal that can be processed by a computer. The continuous signal is sampled at equal intervals according to a preset sampling frequency, and the signal voltage amplitude at each sampling moment is converted into a digital quantity, forming a discrete data sequence arranged in chronological order. Each data point in this sequence corresponds to a specific timestamp and signal amplitude, and also includes core characteristics such as waveform rise time, duration, peak value, and frequency distribution, collectively constituting the feature sequence and providing basic data for subsequent filtering processing.

[0019] Subsequently, an adaptive filtering operation is used to remove noise interference components present in the feature sequence, obtaining the residual signal. Specifically, a least mean square adaptive filtering algorithm is employed. First, background noise under the wind turbine shutdown state is collected as a reference noise sequence. The original feature sequence containing defect signals and noise is used as the main input sequence. The algorithm dynamically adjusts the filtering coefficients by minimizing the mean square error criterion, canceling out components related to the reference noise from the main input sequence, and finally outputting the noise-removed residual signal. The step size factor of the algorithm is determined by combining filtering stability and convergence speed. In this embodiment, the step size factor is set to 0.005-0.01, and this value range is determined through numerous field noise filtering experiments.

[0020] Finally, if the residual signal retains complete signal intensity variation characteristics, it is determined to be a cleaned signal. The cleaned signals are then integrated to obtain a denoised signal dataset. It should be noted that the core of determining the validity of the residual signal is to verify whether the filtering process retains the defect-related acoustic emission characteristics, avoiding distortion and loss of the effective signal during filtering. The determination criteria include five core acoustic emission characteristic parameters: peak value, rise time, duration, energy, and center frequency. By comparing the rate of change of these five parameters before and after filtering, it is determined whether the effective features are completely retained. The preset allowable range of characteristic change rate is determined through a standard lead-breaking test. The lead-breaking test is a common simulated sound source test in the field of acoustic emission. A standard transient acoustic emission signal is generated by breaking a pencil lead. The changes in characteristic parameters of the standard signal before and after filtering are collected, and the maximum allowable change rate that can completely preserve the signal characteristics is obtained by statistical analysis. That is, when the change rate of the five core characteristic parameters of the filtered residual signal is within ±5%, the residual signal is judged to retain the complete signal intensity change characteristics and belongs to the cleaned signal. Non-cleaned signals are eliminated. After the validity of all signal segments is determined, the cleaned signals of all channels are aligned by timestamp. Empty signals with an amplitude of 0 and invalid signals with a duration exceeding the reasonable range of 100μs-1000μs are eliminated. The signals are then structured and integrated according to the sensor channel and time sequence to form a denoised signal dataset containing signal timestamps, sensor numbers, waveform data, and core characteristic parameters.

[0021] In step S12, based on the denoised signal dataset, and for the pre-acquired signal timestamp records and signal acquisition duration, a time-frequency decomposition operation is used to separate specific frequency intervals related to the damage, determining the distribution range of potential damage signals, including: Obtain signal timestamp records and signal acquisition duration. For the denoised signal dataset, divide it into continuous time windows based on the signal timestamp records and signal acquisition duration to obtain segmented signal sequences. The segmented signal sequence is decomposed into time-frequency components using a wavelet packet transform algorithm to obtain a time-frequency energy matrix containing information on multiple bandwidths. Extract the spectral envelope features from the time-frequency energy matrix and calculate the transient characteristic amplitude within each frequency band. If the transient characteristic amplitude is greater than the preset damage energy threshold, the corresponding bandwidth is determined to be a specific frequency range related to the damage, and the distribution range of the potential damage signal is determined according to the continuous state of the specific frequency range within the time window.

[0022] First, signal timestamp records and signal acquisition durations are obtained. For the denoised signal dataset, continuous time windows are divided according to the signal timestamp records and signal acquisition durations to obtain segmented signal sequences. It should be noted that the signal timestamp records contain the precise acquisition time corresponding to each data point in the denoised signal dataset, and the signal acquisition duration is the total acquisition time of this acoustic emission signal. The core of time window division is to divide a long continuous signal sequence into multiple analysis units of equal length, taking into account both time resolution and frequency resolution, ensuring that the transient characteristics of a single defect acoustic emission event can be completely captured. The duration of acoustic emission events generated by typical defects in wind turbine blades is concentrated between 100μs and 1000μs. A non-overlapping fixed time window is adopted, with a window length set to 1024 data points, corresponding to a time length of 512μs at a 2MHz sampling frequency. According to this window length, the continuous signal sequence of the denoised signal dataset is divided into multiple continuous segmented signal sequences. Each segmented signal sequence corresponds to a clear start and end time of the time window, providing standardized analysis units for subsequent time-frequency decomposition.

[0023] Subsequently, the segmented signal sequence was decomposed using a wavelet packet transform algorithm to obtain a time-frequency energy matrix containing information on multiple frequency band widths. It should be noted that the db6 wavelet basis was used, and the number of decomposition layers was determined based on the sampling frequency and the required bandwidth resolution. For a sampling frequency of 2MHz, the number of decomposition layers was set to 6. The db6 wavelet basis has good regularity and tight support, making it suitable for extracting transient features of acoustic emission signals from composite materials. Analysis of the bandwidth energy distribution of different types of defect signals showed that matrix cracking signals were concentrated in the 100-150kHz range, fiber fracture signals in the 200-300kHz range, and interface debonding signals in the 50-100kHz range. The 6-layer decomposition divided the signal at a 2MHz sampling frequency into 64 bandwidths, each with a bandwidth of 15.625kHz. This allowed the typical defect frequency ranges to fall into different bandwidths (corresponding to bandwidths 4-10, 13-19, and 3-7, respectively), avoiding feature confusion caused by bandwidth overlap. After wavelet packet decomposition of each segmented signal sequence, the signal energy at each time point and in each frequency band is calculated. The signal energy is the sum of the squares of the signal amplitudes in that frequency band, ultimately forming a two-dimensional time-frequency energy matrix. The rows of the matrix correspond to the time points within the time window, and the columns correspond to the 64 frequency bands after decomposition. Each element value in the matrix represents the signal energy at the corresponding time point and in the corresponding frequency band, fully characterizing the distribution of signal energy in the time and frequency domains.

[0024] Subsequently, the spectral envelope features of the time-frequency energy matrix are extracted, and the transient characteristic amplitudes within each frequency band are calculated. Specifically, the extraction of the spectral envelope features is achieved through Hilbert transform. A Hilbert transform is performed on the time series of each frequency band in the time-frequency energy matrix to obtain the analytic signal. The amplitude of the analytic signal is the signal envelope. The spectral envelope features reflect the overall contour of signal energy changes with frequency, clearly showing the energy concentration in different frequency bands. The transient characteristic amplitudes are calculated on a frequency band basis, calculating the peak energy, root mean square value, and total energy of each frequency band within the entire time window. In this embodiment, the peak energy within each frequency band is used as the transient characteristic amplitude. This parameter directly reflects the energy release intensity of transient acoustic emission events within that frequency band, exhibiting high sensitivity to transient signals caused by defects.

[0025] Finally, if the transient characteristic amplitude is greater than the preset damage energy threshold, the corresponding bandwidth is determined to be a specific frequency range related to the damage, and the distribution range of the potential damage signal is determined based on the duration of the specific frequency range within the time window. Specifically, the preset damage energy threshold is determined by combining data from two aspects. On the one hand, historical acoustic emission data of the wind turbine under healthy conditions for 30 days of no-load and rated load operation are collected, and the mean and standard deviation of the transient characteristic amplitude in each frequency band are statistically analyzed. Using the 3σ principle, the damage energy threshold of each frequency band is set to the mean amplitude of that frequency band plus three times the standard deviation. On the other hand, through laboratory blade fatigue tests, fatigue loading is applied to blade specimens of the same model, and acoustic emission signals generated by four typical defects—matrix cracking, fiber breakage, interface debonding, and delamination—are collected simultaneously. The minimum transient energy amplitude corresponding to different defect types is statistically analyzed, and this minimum value is used as a supplementary calibration value for the threshold to ensure that the threshold can cover all types of minor defects and avoid missed detections. When the transient characteristic amplitude of a certain frequency band exceeds the damage energy threshold of the corresponding frequency band and persists for three or more consecutive time windows, the frequency band is determined to be a specific frequency range related to damage, avoiding misjudgment caused by a single random interference signal. After determining the specific frequency range, the distribution range of potential damage signals on the time axis is determined based on the start and end times of the time windows in which the frequency range persists. At the same time, combined with the arrival time difference of signals from different channels in the sensor array, the three-dimensional spatial position of the damage source on the blade is calculated using the triangulation method. Finally, a potential damage signal distribution range containing time range, frequency range, and spatial position range is formed, providing an accurate analysis range for subsequent defect feature calibration.

[0026] In step S13, the signal intensity variation characteristics within the distribution range are acquired. These characteristics are calibrated using a pre-acquired acquisition device sensitivity to obtain the signal energy distribution. If the signal energy distribution exceeds a preset energy threshold, it is marked as a suspected defect signal segment, resulting in a suspected defect signal set, including: Obtain the signal intensity variation characteristics within the distribution range and extract the transient amplitude sequence; The signal energy distribution is obtained by performing attenuation compensation processing on the transient amplitude sequence using the pre-acquired sensitivity of the acquisition device. If the signal energy distribution exceeds a preset energy threshold, the corresponding region will be marked as a suspected defective signal segment. The suspected defect signal segments are summarized to obtain a suspected defect signal set.

[0027] First, the signal intensity variation characteristics within the distribution range are obtained, and the transient amplitude sequence is extracted. It should be noted that the signal intensity variation characteristics refer to the changes in signal amplitude over time within the corresponding time window and specific frequency range within the distribution range of the potential damage signal determined in step S12. This directly reflects the energy release process of the acoustic emission event from the defect. The signal envelope amplitude within this distribution range is extracted and arranged in chronological order to form a transient amplitude sequence.

[0028] Subsequently, the transient amplitude sequence is attenuated and compensated using the sensitivity of the acquisition equipment to obtain the signal energy distribution. Specifically, sensor sensitivity calibration is first performed. Each acoustic emission sensor is provided with a calibrated frequency response curve at the factory. Based on the specific frequency range determined in step S12, the sensor's sensitivity value within that frequency range is found. The measured voltage amplitude is converted into the actual particle vibration velocity amplitude, eliminating systematic errors caused by sensitivity differences between different sensors and at different frequencies. Next, acoustic wave propagation attenuation compensation is performed. The acoustic attenuation coefficient of the composite material is obtained through laboratory standard testing. For the material used in this blade, a lead-breaking test is conducted. Sensors are placed at different distances on the specimen to collect standard acoustic emission signals generated by lead breakage. The attenuation law of the signal amplitude with propagation distance is fitted, and the attenuation coefficient of the material within a specific frequency range is calculated. Multiply the propagation distance from the damage source to each sensor by the attenuation coefficient to obtain the signal propagation attenuation. Perform reverse compensation on the amplitude measured by each sensor. After completing sensitivity calibration and attenuation compensation, calculate the signal energy in each time window. The signal energy is the integral of the square of the calibrated amplitude in time within that window. Finally, obtain the signal energy distribution at each time point and each spatial location.

[0029] Subsequently, if the signal energy distribution exceeds a preset energy threshold, the corresponding area is marked as a suspected defect signal segment. Specifically, the preset energy threshold is determined by combining two sets of data. First, through laboratory blade defect simulation tests, different sizes of matrix cracks, fiber fractures, interface debonding, and delamination defects are created on blade specimens of the same model. Fatigue loading tests are then conducted, and calibrated acoustic emission signal energy corresponding to different defect severity levels is collected simultaneously to determine the minimum energy threshold that can reflect the actual defect. Second, the minimum calibrated signal energy corresponding to defects confirmed by shutdown and maintenance in historical defect cases of the same model of wind turbine in the wind farm is statistically analyzed. The smaller of the two minimum values ​​from the laboratory test and the field case is taken and multiplied by a safety margin factor of 0.8 as the final preset energy threshold. When the calibrated signal energy distribution exceeds the preset energy threshold within a certain continuous time region, the signal segment corresponding to that time region is marked as a suspected defect signal segment.

[0030] Finally, the suspected defect signal segments are summarized to obtain a suspected defect signal set. Specifically, all marked suspected defect signal segments are deduplicated, removing duplicate marked segments generated by the same acoustic emission event collected by different sensors. Multiple signal segments generated by the same damage source are merged into an independent acoustic emission event by correlation and matching based on the occurrence time, frequency characteristics, and location of the acoustic emission event. Subsequently, all independent suspected defect acoustic emission events are sorted according to their occurrence time and blade location, forming a structured suspected defect signal set. Each element in the set corresponds to an independent suspected defect event, containing complete time, frequency, spatial, and energy characteristic information.

[0031] like Figure 2 As shown, in step S14, historical vibration data and operating load records for the time interval corresponding to the suspected defect signal set are acquired. Waveform envelope deviation analysis is performed in conjunction with preset historical defect annotations to identify abnormal fluctuation segments and determine the defect signal interval, including: S141, acquire historical vibration data and operating load records for the time interval corresponding to the suspected defect signal set, align the historical vibration data and the operating load records to obtain a multi-dimensional time-series state matrix; S142, compare the multidimensional temporal state matrix with the preset historical defect annotations to obtain a local waveform similarity sequence; S143, extract waveform envelope features based on the local waveform similarity sequence, calculate the deviation between the waveform envelope features and the preset standard waveform envelope, and if the deviation is greater than the preset deviation threshold, it is determined to be an abnormal fluctuation segment. S144, Obtain the start and end time nodes of the abnormal fluctuation segment, and determine the defect signal range by processing the start and end time nodes through feature mapping.

[0032] First, historical vibration data and operational load records for the time interval corresponding to the suspected defect signal set are acquired. These historical vibration data and operational load records are then aligned to obtain a multi-dimensional time-series state matrix. Specifically, the historical vibration data consists of vibration data collected by acceleration vibration sensors on the wind turbine blades during the same time period as the suspected defect signal segment, with a sampling frequency of 1kHz-10kHz. This data includes characteristic parameters such as the effective value, peak value, dominant frequency, and phase of the vibration acceleration. The operational load records are operational parameters collected by the wind turbine's main control system during the same time period, including impeller speed, real-time wind speed, and wind direction angle. These parameters reflect the stress conditions of the blades when the defect occurs. Using the timestamp of the acoustic emission signal as a reference, cubic spline interpolation is used to resample the vibration data, and linear interpolation is used to resample the slowly changing operational load records. This ensures that the temporal resolution of the vibration data and operational load records is completely consistent with the acoustic emission signal, guaranteeing that each time point corresponds to a set of synchronized acoustic emission characteristics, vibration characteristics, and operational load characteristics. The aligned multi-dimensional features are arranged in chronological order to form a multi-dimensional temporal state matrix. The rows of the matrix correspond to the synchronization time points, and the columns correspond to different feature dimensions. In this embodiment, the feature dimensions include seven dimensions: acoustic emission energy, effective value of vibration acceleration, vibration dominant frequency, impeller speed, generator power, blade pitch angle, and wind speed. Each element in the matrix corresponds to a specific value of a certain time point and a certain feature dimension, which fully characterizes the synchronous changes of the blade's signal characteristics, dynamic response, and operating conditions during the defect occurrence process.

[0033] Subsequently, the multidimensional time-series state matrix and the preset historical defect annotations are compared to obtain a local waveform similarity sequence. It should be noted that the preset historical defect annotation library is a standardized defect feature sample library constructed from multi-source data. The sample source is five years of operational data from the same type of wind turbine in the wind farm since its commissioning, containing multidimensional time-series feature samples corresponding to different types and severity of blade defects confirmed through shutdown maintenance. Each sample in the historical defect annotation library contains the same seven feature dimensions as the multidimensional time-series state matrix, and also annotates the corresponding defect type and severity, providing a standard reference for similarity comparison. A one-dimensional convolutional neural network model is used for feature extraction and similarity calculation. This one-dimensional convolutional network can automatically extract the deep features of the multidimensional time-series data and accurately calculate the similarity between the current time-series features and the historical defect annotations. The input to this one-dimensional convolutional neural network model is a multi-dimensional temporal feature vector with a length of 256 and a dimension of 7, which perfectly matches the feature dimension of the multi-dimensional temporal state matrix. The model structure includes an input layer, two one-dimensional convolutional layers, using 32 and 64 convolutional kernels of length 3 respectively, with ReLU activation function. Each convolutional layer is followed by a max pooling layer for dimensionality reduction and extraction of key features, then a flattening layer and two fully connected layers with 128 and 64 neurons respectively, using ReLU activation function. The final output layer uses Sigmoid activation function and outputs a similarity value between 0 and 1. The closer the value is to 1, the higher the similarity between the current temporal feature and the historical defect annotation; the closer it is to 0, the lower the similarity.

[0034] The pre-training process of this convolutional neural network is supervised learning. The training data comes from 5 years of operating data of the same type of wind turbine in the wind farm and the full life cycle data of laboratory blade fatigue tests. First, multi-dimensional time-series feature vectors corresponding to each sample are extracted, and each feature vector is assigned a category label, with 0 representing a healthy state and 1 representing a defective state. At the same time, the corresponding defect type is labeled to construct a sample dataset covering four common defect types: matrix cracking, fiber fracture, interface debonding, and delamination. Then, the dataset is Z-score standardized so that the mean of each feature dimension is 0 and the variance is 1, eliminating the influence of different feature dimensions and numerical ranges on the model. The standardized dataset is randomly divided into training and test sets in a 7:3 ratio. Binary cross-entropy is used as the loss function, and the Adam optimizer is used with an initial learning rate of 0.001 for model training. During the training process, 5-fold cross-validation is used to verify the generalization ability of the model, and early stopping is used to prevent the model from overfitting. The multidimensional temporal state matrix is ​​input into the pre-trained convolutional neural network model through a sliding window. The model outputs the similarity value between each time window and the historical defect annotation, and arranges them in chronological order to form a local waveform similarity sequence.

[0035] Subsequently, waveform envelope features are extracted based on the local waveform similarity sequence, and the deviation between the waveform envelope features and the preset standard waveform envelope is calculated. If the deviation is greater than a preset deviation threshold, it is determined to be an abnormal fluctuation segment. Specifically, the envelope of the local waveform similarity sequence is first extracted using Hilbert transform to obtain waveform envelope features, which reflect the overall trend of similarity over time. Deviation refers to the degree of difference between the current waveform envelope features and the normal similarity sequence envelope features under the healthy state of the wind turbine. The deviation is calculated using Euclidean distance. The normal similarity sequence envelope features under the healthy state of the wind turbine are obtained by collecting 30 days of operational data confirmed to be defect-free during the initial operation of the wind turbine or after a major overhaul, extracting the envelope line through Hilbert transform, and then taking the average value. The preset deviation threshold is obtained by statistically analyzing the rated operational data of the wind turbine under the healthy state for 30 days. Multidimensional time-series data under the healthy state is collected, and its similarity sequence envelope features with healthy samples are calculated. The mean and standard deviation of the deviation are statistically analyzed. Using the 3σ principle, the deviation threshold is set to the mean plus three times the standard deviation. When the deviation of the waveform envelope features continuously exceeds the preset deviation threshold, the time period is judged as an abnormal fluctuation segment. This ensures that the segment is significantly different from the healthy state and highly matched with historical defect features, avoiding misjudgments caused by random interference.

[0036] Finally, the start and end time points of the abnormal fluctuation segments are obtained, and these time points are processed through feature mapping to determine the defect signal interval. Specifically, the start and end times of the abnormal fluctuation segments are first recorded to obtain precise start and end time points. The core of the feature mapping process is to map these start and end time points back to the original acoustic emission signal time axis. Combining the acoustic emission frequency characteristics, spatial positioning characteristics, and energy characteristics within this time period, the time interval is finely corrected, eliminating small-amplitude normal signal segments within the interval. Simultaneously, by combining the historical defect annotations with the highest similarity, the defect type corresponding to this interval is determined. The finally determined defect signal interval includes a precise range in three dimensions: first, the time range, i.e., the precise start and end times of the defect signal; second, the frequency range, i.e., the specific frequency interval corresponding to the defect signal; and third, the spatial range, i.e., the three-dimensional position coordinate interval of the defect on the blade. The defect type corresponding to this interval is also labeled, completing the determination from suspected defects to precise defect intervals.

[0037] In step S15, based on the defect signal range and combined with pre-acquired ambient temperature changes and wind speed fluctuations, a correlation mapping operation is used to calculate the matching quantization value between the defect signal range and the blade stress distribution. The matching quantization value is then weighted to obtain a defect impact assessment index, including: Obtain the ambient temperature sequence and wind speed fluctuation data for the defect signal range; Thermodynamic and aerodynamic coupling processing is performed based on the ambient temperature sequence and the wind speed fluctuation data to obtain the blade stress distribution matrix; The defect signal range is mapped to the blade stress distribution matrix to extract local stress concentration features; By employing correlation mapping, the local stress concentration characteristics are compared with a preset material fatigue damage model to obtain a quantitative value of the matching degree. The quantitative value of the matching degree is then weighted to obtain a defect impact assessment index.

[0038] First, the ambient temperature sequence and wind speed fluctuation data for the defect signal interval are acquired. Specifically, the ambient temperature sequence includes ambient air temperature data collected by the wind farm meteorological station within the corresponding time period of the defect signal interval, as well as temperature data of the windward and leeward sides of the blade collected by the blade surface temperature sensor. The sampling frequency is 1Hz, which reflects the thermal environment and internal temperature gradient of the blade. The wind speed fluctuation data is three-dimensional wind speed data collected by the ultrasonic anemometer on the top of the wind turbine nacelle within the same time period, including wind speed magnitude, wind direction angle, and turbulence intensity. The sampling frequency is 10Hz, which reflects the real-time changes in the aerodynamic load on the blade. These two types of data together determine the external load on the blade when the defect occurs and are the core input parameters for calculating the blade stress distribution.

[0039] Subsequently, thermodynamic and aerodynamic coupling processing is performed based on the ambient temperature sequence and wind speed fluctuation data to obtain the blade stress distribution matrix. Specifically, thermodynamic analysis is first performed. Based on the ambient temperature sequence and blade surface temperature data, combined with the three-dimensional finite element model of the blade provided by the blade manufacturer, and material performance parameters such as thermal conductivity and coefficient of thermal expansion of the blade composite material (these parameters are provided by the material manufacturer), the transient temperature field distribution inside the blade is calculated using the finite element thermal analysis method. If the blade manufacturer does not provide a three-dimensional finite element model of the blade, a three-dimensional geometric model of the blade is created based on the blade structural drawings, and a three-dimensional finite element model of the blade is established based on the finite element method. Then, aerodynamic analysis is performed. Based on wind speed fluctuation data, synchronized impeller speed, and blade pitch angle parameters, the aerodynamic load at each blade element position is calculated using blade element momentum theory, and then the structural stress distribution generated by the aerodynamic load is calculated.

[0040] Blade element momentum theory is a commonly used method in the wind power industry for calculating blade aerodynamic loads. Its core principle is to divide the blade along its span into multiple micro-elements, calculate the aerodynamic load of each element separately, and then integrate along the span to obtain the load and stress distribution of the entire blade. Finally, thermodynamic and aerodynamic coupling is performed, linearly superimposing the thermal and aerodynamic stresses at each moment to obtain the total stress distribution of the blade at each moment and spatial location, forming a blade stress distribution matrix. The dimension of this matrix is ​​time × spatial location × stress value, where the time dimension corresponds to each time point within the defect signal interval, and the spatial location corresponds to the mesh nodes of the blade's three-dimensional finite element model. In this embodiment, the blade is divided into 10,000 mesh nodes. This number of nodes is determined by comprehensively considering the accuracy of transient stress solution and the computational cost of the system. On the one hand, this mesh density ensures that the mesh size reaches the centimeter or even millimeter level in the critical load-bearing parts of the blade (such as the blade root and main beam), meeting the accuracy requirements for sensitively capturing instantaneous stress concentration and stress gradient changes induced by minute defects. On the other hand, when performing three-dimensional transient coupled calculations in the time and spatial domains, the computational workload of the stress distribution matrix generated by 10,000 nodes can be controlled within the real-time or near-real-time processing capabilities of existing industrial-grade embedded controllers, avoiding the curse of dimensionality and computational delays caused by excessive mesh subdivision. Each node corresponds to a three-dimensional coordinate and stress value, completely characterizing the transient stress distribution across the entire blade range during the defect occurrence process.

[0041] Next, the defect signal interval is mapped to the blade stress distribution matrix to extract local stress concentration features. It should be noted that the core of the mapping process is to accurately map the time and spatial range of the defect signal interval to the time and spatial dimensions of the blade stress distribution matrix, extracting stress distribution data within this time-space range for feature extraction. The local stress concentration features include four core parameters: first, the maximum principal stress value in the region, which is the maximum value of the principal stress in the region, reflecting the maximum stress level experienced by the defect area; second, the stress concentration coefficient, obtained by calculating the ratio of the maximum stress in the defect area to the average stress of the blade at the same cross section; third, the stress gradient, obtained by spatial difference calculation of the stress field in the defect area; and fourth, the number of stress cycles, i.e., the number of alternating stress cycles experienced by the region during wind turbine operation, obtained by statistically analyzing the stress time history of the region based on the rainflow counting method.

[0042] Finally, correlation mapping technology is used to compare the local stress concentration characteristics with the preset material fatigue damage model to obtain a quantitative value of the matching degree. This quantitative value of the matching degree is then weighted to obtain a defect impact assessment index. Specifically, correlation mapping technology is used to compare the local stress concentration characteristics with the preset material fatigue damage model to obtain a quantitative value of the matching degree. The preset material fatigue damage model is based on composite material fatigue theory and combined with experimental data of the material used in the blade. The model construction process involves first conducting a standard fatigue test on the glass fiber reinforced epoxy resin composite material used in the blade, testing the fatigue life of the material under different stress levels, and fitting the SN curve (stress-life curve) of the material. Combined with the Miner linear cumulative damage criterion, a cumulative fatigue damage model of the composite material is established. Simultaneously, during the fatigue test, acoustic emission signal characteristics and stress concentration characteristics under different damage levels are collected to establish a mapping relationship between stress concentration characteristics, fatigue damage level, and acoustic emission signal characteristics, forming a complete material fatigue damage model. The model divides fatigue damage into five levels, from level 1 (minor damage) to level 5 (severe damage).

[0043] Grey relational analysis was employed to achieve correlation mapping. Extracted local stress concentration features were used as a reference sequence, and standard stress concentration features corresponding to the five damage levels in the material fatigue damage model were used as comparison sequences. The grey relational degree between the reference sequence and each comparison sequence was calculated. The grey relational degree ranged from 0 to 1; the closer the value was to 1, the higher the match between the current local stress concentration feature and the standard feature of the corresponding damage level, and the more severe the fatigue damage corresponding to the defect. The highest grey relational degree was used as the quantitative value of the matching degree, and the corresponding damage level was determined. Finally, the quantitative values ​​of the matching degree were weighted to obtain the defect impact assessment index.

[0044] The weighting coefficients were determined based on blade structure design specifications such as GB / T25383-2010 "Wind Turbine Blades of Wind Turbine Generators", wind power industry standards, and statistical analysis of failure cases of over 10,000 wind turbine blades in China. Combining the blade structure's force transmission path and defect failure risk level, the coefficients were set according to three dimensions: defect location, service life, and material properties. Specifically, the weighting coefficient was set to 1.5 for defects located in critical load-bearing areas such as the blade root bolt connection area and the main beam cap; 1.3 for defects located in secondary critical areas such as the leading and trailing edge bonding seams; and 1.3 for defects located in non-critical areas such as the blade tip and blade surface. The weighting coefficient is set to 0.8; for blades with accumulated operating time exceeding 70% of their design life, the weighting coefficient is set to 1.3; for those exceeding 50% but less than 70% of their design life, the weighting coefficient is set to 1.2; for those less than 50% of their design life, the weighting coefficient is set to 1.0; and for the remaining strength of the material in the defect area being less than 80% of the design value, the weighting coefficient is set to 1.2. The remaining strength of the material in the defect area is calculated based on the aforementioned material fatigue damage model, using the historical stress cycle count and SN curve of the defect area to determine the cumulative damage degree. The remaining strength is calculated by subtracting the cumulative damage degree and multiplying it by the initial design strength. The quantified matching degree value is multiplied by the corresponding weighting coefficients and then normalized to obtain a defect impact assessment index ranging from 0 to 100 points. A higher score indicates a greater impact of the defect on the blade structure safety and a higher operational risk.

[0045] In step S16, if the defect impact assessment index exceeds a preset safety threshold, a matching analysis is performed based on a preset state response template, combined with pre-acquired cumulative blade operating time and data from the defect signal range, to generate defect state information and determine the defect concern level, including: If the defect impact assessment index exceeds the preset safety threshold, the cumulative operating time of the blade is obtained, and the blade position coordinates corresponding to the defect signal interval data are extracted. The overlapping area with the preset structural weak point distribution is determined based on the blade position coordinates; The overlapping area is used to match the preset state response template. The urgency level identifier in the state response template is extracted and sorted in combination with the cumulative operating time of the blade to determine the defect concern level.

[0046] First, if the defect impact assessment index exceeds a preset safety threshold, the cumulative operating time of the blade is acquired, and the blade position coordinates corresponding to the defect signal interval data are extracted. It should be noted that the preset safety threshold is determined based on wind power industry standards, blade manufacturer structural design requirements, and a large number of wind farm blade operation, maintenance, and failure cases. In this embodiment, the safety threshold for the defect impact assessment index is set to 60 points. When the assessment index exceeds 60 points, the defect is considered to have exceeded the permissible range for safe blade operation, requiring maintenance intervention. Simultaneously, a tiered threshold is set: 60-70 points for low risk, 70-85 points for medium risk, and above 85 points for high risk. Different risk levels correspond to different levels of maintenance urgency. The determination of the safety threshold also incorporates the allowable stress of the blade material. When the maximum stress in the defect area exceeds 80% of the material's allowable stress, the corresponding assessment index is exactly 60 points, ensuring that the threshold setting meets the blade structural safety design requirements and avoiding over-maintenance or under-maintenance. The cumulative operating time of the blades is the total number of operating hours recorded by the wind turbine's main control system since the blades were put into operation, including full-load operating time, partial-load operating time, and number of start-stop cycles. These data reflect the overall fatigue accumulation and material aging status of the blades. When the defect impact assessment index exceeds the safety threshold of 60 points, the precise location coordinates of the defect in the blade's three-dimensional coordinate system are extracted from the defect signal interval data to form a three-dimensional coordinate set of the defect area.

[0047] Next, the overlapping area with the preset structural weak point distribution is determined based on the blade position coordinates. It should be noted that the preset structural weak point distribution is based on blade design drawings, finite element strength analysis results, manufacturing records, and statistical analysis of historical failure cases. High-risk areas on the blade that have a significant impact on structural safety and are prone to defects are specifically divided into the blade root bolt connection area (high stress concentration area), the main beam cap end area (critical load transfer area), the web bonding joint (easily debonded area), and the leading and trailing edge bonding joints (adhesive layer aging area). Each weak point area is marked with a specific coordinate range and risk level in the blade's three-dimensional coordinate system, forming a complete blade structural weak point distribution map. The extracted defect location coordinate set is spatially superimposed with the structural weak point distribution map, and the overlapping area between the defect area and the structural weak point area is calculated.

[0048] Finally, the overlapping area is matched with a preset status response template. The urgency level identifier in the status response template is extracted and sorted according to the cumulative operating time of the blade to determine the defect concern level. Specifically, the preset status response templates are compiled based on industry standards, the blade manufacturer's official maintenance manual, and a large number of field maintenance cases from major domestic wind power operation and maintenance companies. The template library has corresponding standardized maintenance instruction templates for different defect types, overlapping area ratios, risk levels, and defect locations. During the matching process, the template with the highest fit is matched from the template library according to the defect type, overlapping area ratio, risk level, and defect location, and the urgency level identifier in the template is extracted. It should be noted that each status response template has a corresponding urgency level identifier, which is divided into four levels: urgent, high, medium, and low. These directly correspond to the risk level of the defect. High-risk defects with a score of 85 or above correspond to the urgent level, medium-risk defects with a score of 70-85 correspond to the high level, low-risk defects with a score of 60-70 correspond to the medium level, and defects with a score of less than 60 correspond to the low level. Each urgency level indicator corresponds to a fixed urgency coefficient: 1.5 for urgent, 1.2 for high, 1.0 for medium, and 0.8 for low. Simultaneously, an operating time coefficient is set based on the cumulative operating time of the blades: 1.4 for cumulative operating time exceeding 70% of the design life; 1.2 for exceeding 50% but less than 70%; and 1.0 for less than 50%. A priority score is calculated for each maintenance instruction, which is the urgency coefficient multiplied by the operating time coefficient. Instructions are sorted from highest to lowest priority score to determine the blade defect concern level for the entire wind farm.

[0049] In step S17, based on the defect concern level, combined with the historical vibration data and the ambient temperature change, a dynamic monitoring plan is generated to continuously track the defect signal range and obtain long-term defect evolution data, including: Based on the defect concern level, the historical vibration data and the ambient temperature change of the defect signal range are time-series aligned to obtain multidimensional correlation features. If the fluctuation amplitude of the multidimensional correlation feature exceeds a preset amplitude threshold, the sampling frequency is increased within the defect signal range, and a dynamic monitoring plan is generated based on the sampling frequency. The defect signal range is continuously tracked through the dynamic monitoring plan to obtain long-term defect evolution data.

[0050] First, based on the defect concern level, the historical vibration data and ambient temperature changes within the defect signal interval are time-series aligned to obtain multidimensional correlation features. Specifically, the historical vibration data consists of full vibration data collected by blade vibration sensors over three consecutive months prior to the time period corresponding to the defect signal interval, including characteristic parameters such as effective vibration acceleration, peak value, dominant frequency, and phase. The ambient temperature change data comprises daily and seasonal variations of ambient air temperature and blade surface temperature within the same time period. The time-series alignment is performed on a daily basis, using time as the benchmark. Linear interpolation is used to resample the vibration and temperature data, unifying their temporal resolution to one hour, ensuring that each time point corresponds to synchronized vibration and temperature characteristics. The core of multidimensional correlation feature extraction is analyzing the correlation between vibration characteristics and ambient temperature changes, specifically including three core parameters: the Pearson correlation coefficient of the effective vibration acceleration with ambient temperature; the maximum offset of the dominant vibration frequency with temperature changes; and the rate of change of vibration energy with the temperature gradient. These three parameters together constitute the multidimensional correlation features.

[0051] Subsequently, if the fluctuation amplitude of the multidimensional correlation feature exceeds a preset amplitude threshold, the sampling frequency is increased within the defect signal range, and a dynamic monitoring plan is generated based on the sampling frequency. It should be noted that the preset amplitude threshold is obtained through statistical analysis of continuous operating data from the previous year under the blade's healthy state. Multidimensional correlation features under healthy state are collected, and the mean and standard deviation of the fluctuation amplitude of each feature parameter are calculated using the 3σ principle. The dynamic adjustment rule for the sampling frequency is determined based on the fluctuation amplitude determination result. Under normal conditions, the sampling frequency for acoustic emission signals is 2MHz, collected for 1 minute per hour; the sampling frequency for vibration data is 5kHz, collected for 1 minute every 10 minutes; and the sampling frequency for temperature data is 1Hz, continuously collected. When the fluctuation amplitude of any parameter of the multidimensional correlation feature exceeds the preset amplitude threshold, it indicates that the defect is in an unstable state. The sampling frequency is immediately increased: acoustic emission signals are continuously collected at 2MHz, vibration data at 10kHz, and temperature data at 10Hz. If the fluctuation amplitude continues to exceed the threshold for 24 hours, a red alert is immediately triggered and pushed to wind farm maintenance personnel for early intervention and inspection. When the fluctuation amplitude falls below the threshold and remains stable for 72 hours, it indicates that the defect is in a stable state, and the sampling frequency can be gradually reduced to normal. The dynamic monitoring plan is generated by comprehensively considering the defect's concern level, the fluctuation of multi-dimensional correlation features, and the sampling frequency adjustment rules. The plan includes monitoring parameter types, sampling frequencies for each parameter, data collection periods, data storage and backup methods, feature extraction and analysis cycles, anomaly warning thresholds and warning procedures, and emergency response measures. Differentiated monitoring plans are developed for defects with different remediation priorities.

[0052] Finally, the defect signal interval is continuously tracked through the aforementioned dynamic monitoring plan to obtain long-term defect evolution data. Specifically, according to the dynamic monitoring plan, the defect signal interval is continuously tracked and monitored over a long period of time, collecting full-dimensional characteristic data of defects under different times, environmental conditions, and operating loads, including core parameters such as maximum acoustic emission energy, effective value of vibration acceleration, stress concentration factor, and defect impact assessment indicators. This data is then structured and organized in chronological order to form long-term defect evolution data. A Long Short-Term Memory (LSTM) network model is used to predict the future development trend and remaining lifespan of defects based on long-term defect evolution data. The input of the LSTM network model is a time-series feature sequence of length 60, which includes six feature dimensions: maximum acoustic emission energy of defects, effective value of vibration acceleration, stress concentration factor, maximum ambient temperature, maximum wind speed, and daily operating hours over the past 60 days. The model structure includes an input layer, two LSTM layers with 128 and 64 neurons respectively, using the tanh function as the activation function, a dropout layer with an inactivation rate set to 0.2 to prevent overfitting, two fully connected layers with 32 and 16 neurons respectively, using the ReLU function as the activation function, and finally an output layer using a linear activation function to output the predicted value of the defect impact assessment index for the next 30 days, as well as the remaining number of days before the defect reaches the critical safety threshold.

[0053] The pre-training process of the Long Short-Term Memory (LSTM) network model is supervised learning. The training data comes from the full life cycle data of blade defects monitoring of 500 wind turbines in 10 wind farms in China over a total of 8 years, as well as the full life cycle defect evolution data of laboratory blade fatigue tests. First, all raw data are cleaned, outliers and missing values ​​are removed, and the data is organized into a continuous time-series feature sequence according to wind turbine number and defect number. Using historical data every 60 days as input and the actual defect impact assessment index corresponding to the next 30 days as labels, a dataset containing 15,000 samples is constructed. Then, the dataset is subjected to Min-Max normalization, mapping all feature values ​​to the range of 0-1. The normalized dataset is randomly divided into training and test sets in an 8:2 ratio. Mean Squared Error (MSE) is used as the loss function, and the Adam optimizer is used with an initial learning rate of 0.001 for model training. During training, 5-fold cross-validation is used to verify the model's generalization ability, and a learning rate decay strategy and early stopping method are used to prevent overfitting.

[0054] In summary, this invention discloses a method for identifying wind turbine blade defects based on multi-feature fusion, comprising: acquiring acoustic emission signal data of wind turbine blades; performing denoising processing on the acoustic emission signal data using adaptive filtering to obtain a denoised signal dataset; based on the denoised signal dataset, using time-frequency decomposition to separate specific frequency intervals related to damage, and determining the distribution range of potential damage signals, for pre-acquired signal timestamp records and signal acquisition duration; acquiring signal intensity variation characteristics within the distribution range, calibrating the signal intensity variation characteristics using pre-acquired acquisition device sensitivity to obtain signal energy distribution, and marking suspected defect signal segments if the signal energy distribution exceeds a preset energy threshold to obtain a suspected defect signal set; acquiring historical vibration data and operating load data for the time interval corresponding to the suspected defect signal set. The invention records data and performs waveform envelope deviation analysis based on preset historical defect annotations to identify abnormal fluctuation segments and determine defect signal intervals. Based on these intervals and pre-acquired ambient temperature and wind speed fluctuations, a correlation mapping operation is used to calculate the matching quantization value between the defect signal intervals and blade stress distribution. This matching quantization value is then weighted to obtain a defect impact assessment index. If the defect impact assessment index exceeds a preset safety threshold, a matching analysis is performed based on a preset state response template, combined with pre-acquired blade operating cumulative time and the data from the defect signal intervals, to generate defect state information and determine the defect concern level. Based on the defect concern level, combined with historical vibration data and ambient temperature changes, a dynamic monitoring plan is generated to continuously track the defect signal intervals and obtain long-term defect evolution data. This invention achieves high-precision identification of wind turbine blade defects, deep fusion of multi-source features, and intelligent maintenance decision support based on dynamic monitoring through defect identification.

[0055] Reference Figure 3 The second embodiment of the present invention provides a wind turbine blade defect identification system based on multi-feature fusion, comprising: The signal acquisition and denoising module is used to acquire acoustic emission signal data of the wind turbine blades and perform denoising processing on the acoustic emission signal data using adaptive filtering to obtain a denoised signal dataset. The time-frequency decomposition module is used to separate specific frequency ranges related to damage based on the denoised signal dataset and the pre-acquired signal timestamp records and signal acquisition duration, thereby determining the distribution range of potential damage signals. The suspected defect marking module is used to acquire the signal intensity change characteristics within the distribution range, calibrate the signal intensity change characteristics using the pre-acquired acquisition device sensitivity, and obtain the signal energy distribution. If the signal energy distribution exceeds a preset energy threshold, it is marked as a suspected defect signal segment, thus obtaining a suspected defect signal set. The defect interval determination module is used to acquire historical vibration data and operating load records of the time interval corresponding to the suspected defect signal set, and perform waveform envelope deviation analysis in combination with preset historical defect labels to identify abnormal fluctuation segments and determine the defect signal interval. The defect impact assessment module is used to calculate the matching quantization value between the defect signal range and the blade stress distribution based on the defect signal range and the pre-acquired ambient temperature change and wind speed fluctuation, and to perform a weighted calculation on the matching quantization value to obtain the defect impact assessment index. The defect concern generation module is used to generate defect status information and determine the defect concern level if the defect impact assessment index exceeds a preset safety threshold, based on a preset status response template and in combination with the pre-acquired cumulative blade operation time and the defect signal range data. The dynamic monitoring module is used to generate a dynamic monitoring plan based on the defect concern level, combined with the historical vibration data and the changes in ambient temperature, to continuously track the defect signal range and obtain long-term defect evolution data.

[0056] It should be noted that the wind turbine blade defect identification system based on multi-feature fusion provided in this embodiment of the invention is used to execute all the process steps of the wind turbine blade defect identification method based on multi-feature fusion in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0057] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0058] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A fan blade defect identification method based on multi-feature fusion, characterized in that, include: Acoustic emission signal data of wind turbine blades are collected, and the acoustic emission signal data is denoised by adaptive filtering to obtain a denoised signal dataset. Based on the denoised signal dataset, time-frequency decomposition is used to separate specific frequency ranges related to damage, based on the pre-acquired signal timestamp records and signal acquisition duration, to determine the distribution range of potential damage signals. The signal intensity variation characteristics within the distribution range are obtained, and the signal intensity variation characteristics are calibrated using the pre-acquired acquisition device sensitivity to obtain the signal energy distribution. If the signal energy distribution exceeds a preset energy threshold, it is marked as a suspected defect signal segment, thus obtaining a set of suspected defect signals. Historical vibration data and operating load records for the time interval corresponding to the suspected defect signal set are obtained, and waveform envelope deviation analysis is performed in combination with preset historical defect labels to identify abnormal fluctuation segments and determine the defect signal interval. Based on the defect signal range, combined with the pre-acquired ambient temperature changes and wind speed fluctuations, a correlation mapping operation is used to calculate the matching quantization value between the defect signal range and the blade stress distribution. The matching quantization value is then weighted to obtain the defect impact assessment index. If the defect impact assessment index exceeds the preset safety threshold, the defect status information is generated and the defect concern level is determined by matching and analyzing the preset status response template with the pre-acquired cumulative blade operation time and the defect signal range data. Based on the defect concern level, combined with the historical vibration data and the changes in ambient temperature, a dynamic monitoring plan is generated to continuously track the defect signal range and obtain long-term defect evolution data.

2. The fan blade defect recognition method based on multi-feature fusion according to claim 1, characterized in that, The acoustic emission signal data of the collected wind turbine blades is denoised using an adaptive filtering operation to obtain a denoised signal dataset, including: Acoustic emission signals generated by the wind turbine blades during operation are acquired through a sensor array; The acoustic emission signal is discretized according to the sampling frequency to obtain a feature sequence containing waveform characteristics and frequency distribution; Adaptive filtering is used to remove noise interference components present in the feature sequence to obtain the residual signal; If the residual signal retains the complete signal intensity change characteristics, then the residual signal is determined to be a cleaned signal; if it does not retain them, then the residual signal is removed, and the cleaned signal is integrated to obtain a denoised signal dataset.

3. The method for fan blade defect recognition based on multi-feature fusion according to claim 1, characterized in that, The step of separating specific frequency intervals related to damage based on the denoised signal dataset and the pre-acquired signal timestamp records and signal acquisition duration using time-frequency decomposition to determine the distribution range of potential damage signals includes: Obtain signal timestamp records and signal acquisition duration. For the denoised signal dataset, divide it into continuous time windows based on the signal timestamp records and signal acquisition duration to obtain segmented signal sequences. The segmented signal sequence is decomposed into time-frequency components using a wavelet packet transform algorithm to obtain a time-frequency energy matrix containing information on multiple bandwidths. Extract the spectral envelope features from the time-frequency energy matrix and calculate the transient characteristic amplitude within each frequency band. If the transient characteristic amplitude is greater than the preset damage energy threshold, the corresponding bandwidth is determined to be a specific frequency range related to the damage, and the distribution range of the potential damage signal is determined according to the continuous state of the specific frequency range within the time window.

4. The wind turbine blade defect identification method based on multi-feature fusion according to claim 1, characterized in that, The process involves acquiring signal intensity variation characteristics within the distribution range, calibrating these characteristics using pre-acquired acquisition device sensitivity, and obtaining a signal energy distribution. If the signal energy distribution exceeds a preset energy threshold, it is marked as a suspected defective signal segment, resulting in a suspected defective signal set, including: Obtain the signal intensity variation characteristics within the distribution range and extract the transient amplitude sequence; The signal energy distribution is obtained by performing attenuation compensation processing on the transient amplitude sequence using the pre-acquired sensitivity of the acquisition device. If the signal energy distribution exceeds a preset energy threshold, the corresponding region will be marked as a suspected defective signal segment. The suspected defect signal segments are summarized to obtain a suspected defect signal set.

5. The wind turbine blade defect identification method based on multi-feature fusion according to claim 1, characterized in that, The process of acquiring historical vibration data and operational load records for the time interval corresponding to the suspected defect signal set, performing waveform envelope deviation analysis in conjunction with preset historical defect annotations, identifying abnormal fluctuation segments, and determining the defect signal interval includes: Acquire historical vibration data and operational load records for the time interval corresponding to the suspected defect signal set, align the historical vibration data and the operational load records, and obtain a multi-dimensional time-series state matrix. By comparing the multidimensional temporal state matrix with the preset historical defect annotations, a local waveform similarity sequence is obtained; Based on the local waveform similarity sequence, waveform envelope features are extracted, and the deviation between the waveform envelope features and the preset standard waveform envelope is calculated. If the deviation is greater than the preset deviation threshold, it is determined to be an abnormal fluctuation segment. The start and end time nodes of the abnormal fluctuation segment are obtained, and the start and end time nodes are processed by feature mapping to determine the defect signal range.

6. The wind turbine blade defect identification method based on multi-feature fusion according to claim 1, characterized in that, Based on the defect signal range and combined with pre-acquired environmental temperature changes and wind speed fluctuations, a correlation mapping operation is used to calculate the matching quantization value between the defect signal range and the blade stress distribution. The matching quantization value is then weighted to obtain defect impact assessment indicators, including: Obtain the ambient temperature sequence and wind speed fluctuation data for the defect signal range; Thermodynamic and aerodynamic coupling processing is performed based on the ambient temperature sequence and the wind speed fluctuation data to obtain the blade stress distribution matrix; The defect signal range is mapped to the blade stress distribution matrix to extract local stress concentration features; By employing correlation mapping, the local stress concentration characteristics are compared with a preset material fatigue damage model to obtain a quantitative value of the matching degree. The quantitative value of the matching degree is then weighted to obtain a defect impact assessment index.

7. The wind turbine blade defect identification method based on multi-feature fusion according to claim 1, characterized in that, If the defect impact assessment index exceeds a preset safety threshold, a matching analysis is performed based on a preset state response template, combined with pre-acquired blade operation cumulative time and data from the defect signal range, to generate defect state information and determine the defect concern level, including: If the defect impact assessment index exceeds the preset safety threshold, the cumulative operating time of the blade is obtained, and the blade position coordinates corresponding to the defect signal interval data are extracted. The overlapping area with the preset structural weak point distribution is determined based on the blade position coordinates; The overlapping area is used to match the preset state response template. The urgency level identifier in the state response template is extracted and sorted in combination with the cumulative operating time of the blade to determine the defect concern level.

8. The wind turbine blade defect identification method based on multi-feature fusion according to claim 1, characterized in that, Based on the defect concern level, combined with the historical vibration data and the environmental temperature changes, a dynamic monitoring plan is generated to continuously track the defect signal range and obtain long-term defect evolution data, including: Based on the defect concern level, the historical vibration data and the ambient temperature change of the defect signal range are time-series aligned to obtain multidimensional correlation features. If the fluctuation amplitude of the multidimensional correlation feature exceeds a preset amplitude threshold, the sampling frequency is increased within the defect signal range, and a dynamic monitoring plan is generated based on the sampling frequency. The defect signal range is continuously tracked through the dynamic monitoring plan to obtain long-term defect evolution data.

9. A wind turbine blade defect identification system based on multi-feature fusion, characterized in that, include: The signal acquisition and denoising module is used to acquire acoustic emission signal data of the wind turbine blades and perform denoising processing on the acoustic emission signal data using adaptive filtering to obtain a denoised signal dataset. The time-frequency decomposition module is used to separate specific frequency ranges related to damage based on the denoised signal dataset and the pre-acquired signal timestamp records and signal acquisition duration, thereby determining the distribution range of potential damage signals. The suspected defect marking module is used to acquire the signal intensity change characteristics within the distribution range, calibrate the signal intensity change characteristics using the pre-acquired acquisition device sensitivity, and obtain the signal energy distribution. If the signal energy distribution exceeds a preset energy threshold, it is marked as a suspected defect signal segment, thus obtaining a suspected defect signal set. The defect interval determination module is used to acquire historical vibration data and operating load records of the time interval corresponding to the suspected defect signal set, and perform waveform envelope deviation analysis in combination with preset historical defect labels to identify abnormal fluctuation segments and determine the defect signal interval. The defect impact assessment module is used to calculate the matching quantization value between the defect signal range and the blade stress distribution based on the defect signal range and the pre-acquired ambient temperature change and wind speed fluctuation, and to perform a weighted calculation on the matching quantization value to obtain the defect impact assessment index. The defect concern generation module is used to generate defect status information and determine the defect concern level if the defect impact assessment index exceeds a preset safety threshold, based on a preset status response template and in combination with the pre-acquired cumulative blade operation time and the defect signal range data. The dynamic monitoring module is used to generate a dynamic monitoring plan based on the defect concern level, combined with the historical vibration data and the changes in ambient temperature, to continuously track the defect signal range and obtain long-term defect evolution data.

Images